基于视频的运动人体行为识别

发布时间：2018-03-06 19:00

本文选题：运动显著区域　切入点：低秩矩阵分解　出处：《中北大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着计算机与互联网的发展,基于计算机视觉的人体行为识别在人类的生产、生活等实际应用中扮演着重要的角色,并取得了长足的发展。目前,国内外在该研究领域主要针对特征提取和分类器设计方面进行创新和改进,尤其在特征提取方面,包括全局特征(如运动能量图,运动历史图等)、局部特征(如Harris-3D、Hessian、HOG/HOF等)。最近出现的基于轨迹特征的行为识别方法,其在光流场中提取特征点,并对特征点进行跟踪与轨迹描述以识别人体行为,该方法识别效果较好,但由于背景和运动相关区域难以区分,尤其是相机存在相对运动的情况下,导致运动特征和轨迹表达有效性不足的问题,本文在上述基础上主要做了以下工作:(1)在显著性轨迹特征提取中,为保证视频中运动区域的时间相关性,将视频在时间方向划分为互不重叠的视频段;其次,为保证其空间的相关性,将视频分为互不重叠视频块,利用光流信息构建视频的运动矩阵,针对出现摄像机相对运动的视频中,由于运动相关区域通常具有无规律与不规则性,通过实现低秩矩阵分解,将原运动矩阵分为低秩部分和稀疏残差部分,从而获取运动显著区域。(2)在复杂场景视频中,基于上述方法检测的运动相关区域中会包含视频背景中的运动显著区域。为去除背景区域像素点,提出了通过边缘检测与背景差分法结合获取完整的运动目标模板,然后结合上述运动显著相关区域获取运动目标显著区域。(3)在跟踪中,采用中值平滑滤波器去除极端值,然后在光流场中采样并判断是否位于显著区域从而提取显著特征点,最后通过类似于现有的稠密轨迹跟踪方法完成迭代跟踪。
[Abstract]:With the development of computer and Internet, human behavior recognition based on computer vision plays an important role in human production, life and other practical applications, and has made great progress. In the field of feature extraction and classifier design at home and abroad, innovation and improvement are carried out, especially in feature extraction, including global features (such as moving energy map). Motion history maps and local features (such as Harris-3D Hessianhe Hessian / Hog / HOF, etc.). Recently developed behavior recognition methods based on trajectory features, which extract feature points in optical flow field, and track and describe feature points to identify human behavior. The method has good recognition effect, but it is difficult to distinguish the background from the motion related areas, especially in the case of relative motion of the camera, which leads to the lack of effectiveness of the expression of motion features and trajectory. Based on the above work, this paper mainly does the following work: 1) in the feature extraction of salient trajectory, in order to ensure the temporal correlation of the moving region in the video, the video is divided into non-overlapping video segments in the time direction; secondly, in order to ensure the temporal correlation of the moving region in the video, the video is divided into non-overlapping video segments. In order to ensure the spatial correlation of the video, the video is divided into non-overlapping video blocks, and the motion matrix of the video is constructed by using the optical flow information. In view of the relative motion of the video camera, the motion related region is usually irregular and irregular. By implementing the low rank matrix decomposition, the original motion matrix is divided into low rank part and sparse residual part, and the motion significant area. The motion-related regions based on the above methods will contain the moving salient regions in the video background. In order to remove the pixels in the background region, a complete template of moving objects is obtained by combining edge detection with background differential method. In tracking, median smoothing filter is used to remove the extreme value, and then sampling and judging whether the salient region is located in the optical flow field to extract the salient feature points. In the end, iterative tracking is accomplished by using dense trajectory tracking methods similar to the existing ones.
【学位授予单位】：中北大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】